BrowseComp
BrowseComp tests whether a browsing agent can find difficult, fact-seeking answers across the web. Its 1,266 problems use short answers that are easier to grade than an open-ended research report.
Top models on BrowseComp — October 15, 2026
As of October 15, 2026, Atria Dawn Preview leads the BrowseComp leaderboard with 92.5% , followed by GPT-5.6 Sol (92.2%) and Ling 3.1 Flash (91.7%).
Atria Dawn Preview
Shanghai Artificial Intelligence Laboratory
GPT-5.6 Sol
OpenAI
Ling 3.1 Flash
InclusionAI
53 modelsAgentic8% of Agentic reference weightCurrentUpdated October 15, 2026
API rates are catalog prices per million input/output tokens, not measured benchmark-run costs. Context is the published model limit. License names come from recorded publisher terms; unrecorded values remain unavailable.
Leaderboard (53 models)
| Rank | Model / configuration | Score | Parameters (B) | Open / closed | API input / output ($/1M) | Context | License |
|---|---|---|---|---|---|---|---|
| 1 | Atria Dawn PreviewShanghai Artificial Intelligence Laboratory | 92.5% | Not reported | Open | Not recorded | 256K | Not recorded |
| 2 | GPT-5.6 SolOpenAI | 92.2% | Not reported | Closed | $4 / $20 | 1.05M | Proprietary |
| 3 | Ling 3.1 FlashInclusionAI | 91.7% | Not reported | Open | Not recorded | 262K | Not recorded |
| 4 | GPT-6 AstraOpenAI | 91.5% | Not reported | Closed | $10 / $50 | 1.05M | Proprietary |
| 5 | Kimi K3Moonshot AI | 91.2% | Not reported | Pending | $3 / $15 | 1.05M | Not recorded |
| 6 | Claude Opus 5Anthropic | 90.8% | Not reported | Closed | $5 / $25 | 1M | Proprietary |
| 7 | GPT-5.5 ProOpenAI | 90.1% | Not reported | Closed | $30 / $180 | 1M | Proprietary |
| 8 | GPT-5.4 ProOpenAI | 89.3% | Not reported | Closed | $30 / $180 | 1.05M | Proprietary |
| 9 | Step 5 PreviewStepFun | 88.7% | Not reported | Open | $1 / $2.7 | 1M | Not recorded |
| 10 | Claude Mythos 5Anthropic | 88% | Not reported | Closed | $10 / $50 | 1M+ | Proprietary |
| 11 | GPT-5.6 TerraOpenAI | 87.5% | Not reported | Closed | $2 / $12 | 1.05M | Proprietary |
| 12 | Ornith-1.5-397BOrnith AI | 86.6% | Not reported | Open | Not recorded | 262K | Not recorded |
| 13 | Claude Sonnet 5Anthropic | 84.7% | Not reported | Closed | $2 / $10 | 1M | Proprietary |
| 14 | GPT-5.5OpenAI | 84.4% | Not reported | Closed | $5 / $30 | 1M | Proprietary |
| 15 | Claude Opus 4.8Anthropic | 84.3% | Not reported | Closed | $5 / $25 | 1M | Proprietary |
| 16 | Claude Opus 4.6Anthropic | 83.7% | Not reported | Closed | $5 / $25 | 1M | Proprietary |
| 17 | MiniMax M3MiniMax | 83.5% | Not reported | Open | $0.3 / $1.2 | 1M | Not recorded |
| 18 | DeepSeek V4 Pro 0813DeepSeek | 83.4% | Not reported | Open | $1.32 / $3.96 | 1M | Not recorded |
| 19 | GPT-5.6 LunaOpenAI | 83.3% | Not reported | Closed | $0.2 / $1.2 | 1.05M | Proprietary |
| 20 | dots3-note PreviewDots Studio | 83.3% | Not reported | Open | Not recorded | 512K | Not recorded |
| 21 | Kimi K2.6Moonshot AI | 83.2% | Not reported | Open | $0.95 / $4 | 256K | Modified MIT |
| 22 | GPT-5.4OpenAI | 82.7% | Not reported | Closed | $2.5 / $15 | 1.05M | Proprietary |
| 23 | DeepSeek V4 Pro (High)DeepSeek | 80.4% | Not reported | Open | Not recorded | 1M | Not recorded |
| 24 | Claude Opus 4.7 (Adaptive)Anthropic | 79.3% | Not reported | Closed | $5 / $25 | 1M | Proprietary |
| 25 | Nemotron 3.5 Super VL 120B A12BNVIDIA | 78% | Not reported | Open | Not recorded | 1M | Not recorded |
| 26 | Nemotron 3.5 Super VL 120B A12B NVFP4NVIDIA | 77.5% | Not reported | Open | Not recorded | 1M | Not recorded |
| 27 | Inkling-SmallThinking Machines Lab | 77.4% | Not reported | Open | $0.58 / $1.44 | 1M | Not recorded |
| 28 | BeamReflection AI | 77.4% | Not reported | Pending | Not recorded | Not recorded | Not recorded |
| 29 | InklingThinking Machines Lab | 77.1% | Not reported | Open | $1.87 / $4.68 | 1M | Not recorded |
| 30 | Step 3.7 FlashStepFun | 75.8% | Not reported | Open | $0.2 / $1.15 | 256K | Not recorded |
| 31 | Agents-A1InternScience | 75.5% | Not reported | Open | Not recorded | 262K | Not recorded |
| 32 | DeepSeek V4 Flash 0731DeepSeek | 73.2% | Not reported | Open | $0.14 / $0.28 | 1M | Not recorded |
| 33 | Ling 3.0 FlashInclusionAI | 72.2% | Not reported | Open | Not recorded | 262K | Not recorded |
| 34 | Qwen3.5 397BAlibaba | 69% | Not reported | Open | $0.6 / $3.6 | 128K | Not recorded |
| 35 | GLM-5.1Z.AI | 68% | Not reported | Open | $1.4 / $4.4 | 203K | MIT |
| 36 | Ornith-1.5-35B-A3BOrnith AI | 67.6% | Not reported | Open | Not recorded | 262K | Not recorded |
| 37 | Agents-A1-4BInternScience | 66.8% | Not reported | Open | Not recorded | 262K | Not recorded |
| 38 | GPT-5.2OpenAI | 65.8% | Not reported | Closed | $1.75 / $14 | 400K | Proprietary |
| 39 | Qwen3.5-122B-A10BAlibaba | 63.8% | Not reported | Open | Not recorded | 262K | Not recorded |
| 40 | GLM-5Z.AI | 62% | Not reported | Open | $1 / $3.2 | 200K | Not recorded |
| 41 | Qwen3.5-27BAlibaba | 61% | Not reported | Open | Not recorded | 262K | Apache-2.0 |
| 42 | Qwen3.5-35B-A3BAlibaba | 61% | Not reported | Open | Not recorded | 262K | Apache-2.0 |
| 43 | Kimi K2.5 (Reasoning)Moonshot AI | 60.6% | Not reported | Closed | $0.6 / $3 | 128K | Proprietary |
| 44 | Kimi K2.5Moonshot AI | 60.6% | Not reported | Open | $0.6 / $3 | 256K | Moonshot |
| 45 | Ornith-1.5-9BOrnith AI | 56.4% | Not reported | Open | Not recorded | 262K | Not recorded |
| 46 | DeepSeek V4 Flash (High)DeepSeek | 53.5% | Not reported | Open | Not recorded | 1M | Not recorded |
| 47 | GLM-4.7Z.AI | 52% | Not reported | Open | Not recorded | 200K | Not recorded |
| 48 | Sarvam 105BSarvam | 49.5% | Not reported | Open | Not recorded | 128K | Not recorded |
| 49 | Solar Pro 4Upstage | 49.2% | Not reported | Closed | $0.3 / $1.2 | 512K | Proprietary |
| 50 | LongCat-Flash-Lite-SparseMeituan | 48.6% | Not reported | Open | Not recorded | 1M | MIT |
| 51 | Nemotron 3 UltraNVIDIA | 44.4% | Not reported | Open | Not recorded | 1M | Not recorded |
| 52 | Nemotron 3.5 Lightning 30B A3B NVFP4NVIDIA | 36.8% | Not reported | Open | Not recorded | 1M | Not recorded |
| 53 | Sarvam 30BSarvam | 35.5% | Not reported | Open | Not recorded | 64K | Apache-2.0 |
BrowseComp (10-agent, prerelease)
Sections 8.11.1 and 8.11.4 report 93.6% for the 10-agent team. The card says these runs used a pre-release configuration without safeguards and are useful for relative, not absolute, comparison; BenchLM therefore keeps the row display-only and explicitly labeled.
| Source rank | Source model or configuration | Score | Source setup | Parameters (B) | Open / closed | API input / output ($/1M) | Context | License |
|---|---|---|---|---|---|---|---|---|
| 1 | Claude Opus 5Anthropic | 93.6% | AnthropicOriginal result | Not reported | Closed | $5 / $25 | 1M | Proprietary |
How to read this leaderboard
BrowseComp is OpenAI’s 1,266-question browsing benchmark. Solving a question can require repeated searches and gathering clues from several pages. The short-answer format makes grading easier but narrows what the result measures.
Variant check: Match the browser tools, model configuration, search budget, and number of attempts before comparing rows. BrowseComp and BrowseComp-Plus use different evaluation environments; their percentages belong in separate tables.
Use BrowseComp as evidence about difficult web retrieval. It does not establish how well an agent writes a cited report, checks every source, or handles the distribution of questions in your product. Test those jobs separately.
Operator receipt: 53 sourced rows are currently displayable on this page; the leading published row is Atria Dawn Preview at 92.5%.
Honest limit: OpenAI notes that the relationship between short-answer BrowseComp performance and real users’ open-ended questions is unclear. Multi-attempt results and single attempts are not interchangeable. API list prices are not measured evaluation costs.
According to BenchLM.ai, Atria Dawn Preview leads the BrowseComp benchmark with a score of 92.5%, followed by GPT-5.6 Sol (92.2%) and Ling 3.1 Flash (91.7%). The top models are clustered within 0.8 points, suggesting this benchmark is nearing saturation for frontier models.
53 models have been evaluated on BrowseComp. The benchmark falls in the Agentic category. BenchAlign v5.8 gives BrowseComp 8% of the Agentic reference weight, so it moves the Agentic leaderboard and the overall ranking. Reference weights are relative weights in the calibrated model, not fixed shares of a score.
About BrowseComp
Year
2025
Tasks
Research questions requiring browsing
Format
Web search and evidence synthesis
Difficulty
Hard web research
BrowseComp is designed to measure real web research behavior, not just latent world knowledge. It rewards models that can plan searches, inspect multiple pages, and avoid shallow answer synthesis.
Freshness and provenance
Version
BrowseComp 2026
Refresh cadence
Quarterly
Staleness state
Current
Question availability
Public benchmark set
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
What does BrowseComp measure?
BrowseComp tests whether a browsing agent can find difficult, fact-seeking answers across the web. Its 1,266 problems use short answers that are easier to grade than an open-ended research report.
Which model scores highest on BrowseComp?
Atria Dawn Preview by Shanghai Artificial Intelligence Laboratory currently leads with a score of 92.5% on BrowseComp.
How many models are evaluated on BrowseComp?
53 AI models have published results on BrowseComp in the BenchLM catalog.
Compare top models on BrowseComp
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 5,500+ readers.
One email each week. Unsubscribe anytime.